Name: urbansim_templates
Owner: Urban Data Science Toolkit
Description: UrbanSim extension for managing model steps
Created: 2018-02-14 23:07:51.0
Updated: 2018-05-15 23:57:43.0
Pushed: 2018-05-21 20:34:33.0
Size: 93
Language: Python
GitHub Committers
| User | Most Recent Commit | # Commits |
|---|
Other Committers
| User | Most Recent Commit | # Commits |
|---|
ModelManager is a small extension to UrbanSim that treats model steps as objects you can interact with, rather than just functions with inputs and outputs. This enables some nice workflows, such as registration of model steps without needing to add them to a models.py file:
urbansim_templates import modelmanager as mm
urbansim_templates.models import OLSRegressionStep
l = OLSRegressionStep('name', parameters)
l.fit()
l.register()
e model step is now registered with Orca and saved to disk. ModelManager tracks and
ads all previously saved steps, just as if they were listed in models.py.
.run(['name'])
needed, the previously saved steps can be re-loaded into class instances that can be
amined, copied, and edited.
l2 = mm.get_model('name')
Prior to ModelManager, the equivalent workflow would be to (1) estimate a model, (2) generate a config file for it, (3) write a function in models.py to register it with Orca, and (4) edit it by modifying the config file or creating a replacement.
ModelManager works directly with the current versions of UrbanSim and Orca, and is fully interoperable with existing UrbanSim projects and model steps.
The goal with this is to make it easier to develop UrbanSim templates – combinations of data schemas and model steps that we can use in multiple projects without requiring too much customization.
The ModelManager workflow will streamline rollout of a template: once the data is in place, we can estimate and validate model steps by walking through pre-filled Jupyter Notebooks.
There are side benefits as well: easier experimentation with model specifications, easier benchmarking and comparison of alternative model flows (including future hooks for auto-specification), and hopefully easier interoperability with GUI components.
The tradeoff is that model steps from a template class will not be as flexible as general-purpose Orca steps. But since there is a lot of commonality among UrbanSim implementations, it should get us pretty far.
To work with ModelManager, a model step class needs to implement the following features:
Ability to save itself to a dictionary (a to_dict() method)
Ability to rebuild itself from a dictionary (a from_dict() class method)
(ModelManager will handle conversion of the dictionaries to/from YAML or some other future storage format.)
Ability to run itself (e.g. a run() method)
Ability to register itself with ModelManager (a register() method) – just a single line of code, implemented by combining the prior three capabilities
If you look in the models directory you'll see an OLS model step class called OLSRegressionStep that implements these features on top of an existing UrbanSim regression model class.
The modelmanager.py file handles centralized saving and loading of the configs.
In an UrbanSim project directory, the steps saved by ModelManager will show up in a file called modelmanager_configs.yaml.
clone https://github.com/udst/urbansim_templates.git
rbansim_templates
on setup.py develop
Demo notebook: Demo.ipynb
Choose a directory to work in, ideally an existing UrbanSim project since ModelManager cannot yet load data into Orca on its own.
For building and registering model steps, use:
urbansim_templates.models import OLSRegressionStep
For loading or inspecting previously saved model steps, use:
urbansim_templates import modelmanager as mm
For example, adding import urbansim_templates to the top of a file like simulation.py will automatically give Orca access to all previously saved steps in the same project directory.
Please refer to the Python files for details of the current API.
Let's keep the master branch clean and runnable. To make fixes or add features, create a new branch, and open a pull request when you're ready to merge code into master. Have someone else review it before merging, if possible.
Here's a rough development wish list. See issues for more details, and feel free to claim an issue or open new ones.
datamanager.py to support loading data and generating computed columns using a similar workflow (Sam M)